1 // SPDX-License-Identifier: GPL-2.0
3 * Memory Migration functionality - linux/mm/migrate.c
5 * Copyright (C) 2006 Silicon Graphics, Inc., Christoph Lameter
7 * Page migration was first developed in the context of the memory hotplug
8 * project. The main authors of the migration code are:
10 * IWAMOTO Toshihiro <iwamoto@valinux.co.jp>
11 * Hirokazu Takahashi <taka@valinux.co.jp>
12 * Dave Hansen <haveblue@us.ibm.com>
16 #include <linux/migrate.h>
17 #include <linux/export.h>
18 #include <linux/swap.h>
19 #include <linux/swapops.h>
20 #include <linux/pagemap.h>
21 #include <linux/buffer_head.h>
22 #include <linux/mm_inline.h>
23 #include <linux/nsproxy.h>
24 #include <linux/pagevec.h>
25 #include <linux/ksm.h>
26 #include <linux/rmap.h>
27 #include <linux/topology.h>
28 #include <linux/cpu.h>
29 #include <linux/cpuset.h>
30 #include <linux/writeback.h>
31 #include <linux/mempolicy.h>
32 #include <linux/vmalloc.h>
33 #include <linux/security.h>
34 #include <linux/backing-dev.h>
35 #include <linux/compaction.h>
36 #include <linux/syscalls.h>
37 #include <linux/compat.h>
38 #include <linux/hugetlb.h>
39 #include <linux/hugetlb_cgroup.h>
40 #include <linux/gfp.h>
41 #include <linux/pfn_t.h>
42 #include <linux/memremap.h>
43 #include <linux/userfaultfd_k.h>
44 #include <linux/balloon_compaction.h>
45 #include <linux/page_idle.h>
46 #include <linux/page_owner.h>
47 #include <linux/sched/mm.h>
48 #include <linux/ptrace.h>
49 #include <linux/oom.h>
50 #include <linux/memory.h>
51 #include <linux/random.h>
52 #include <linux/sched/sysctl.h>
54 #include <asm/tlbflush.h>
56 #include <trace/events/migrate.h>
60 int isolate_movable_page(struct page
*page
, isolate_mode_t mode
)
62 const struct movable_operations
*mops
;
65 * Avoid burning cycles with pages that are yet under __free_pages(),
66 * or just got freed under us.
68 * In case we 'win' a race for a movable page being freed under us and
69 * raise its refcount preventing __free_pages() from doing its job
70 * the put_page() at the end of this block will take care of
71 * release this page, thus avoiding a nasty leakage.
73 if (unlikely(!get_page_unless_zero(page
)))
77 * Check PageMovable before holding a PG_lock because page's owner
78 * assumes anybody doesn't touch PG_lock of newly allocated page
79 * so unconditionally grabbing the lock ruins page's owner side.
81 if (unlikely(!__PageMovable(page
)))
84 * As movable pages are not isolated from LRU lists, concurrent
85 * compaction threads can race against page migration functions
86 * as well as race against the releasing a page.
88 * In order to avoid having an already isolated movable page
89 * being (wrongly) re-isolated while it is under migration,
90 * or to avoid attempting to isolate pages being released,
91 * lets be sure we have the page lock
92 * before proceeding with the movable page isolation steps.
94 if (unlikely(!trylock_page(page
)))
97 if (!PageMovable(page
) || PageIsolated(page
))
100 mops
= page_movable_ops(page
);
101 VM_BUG_ON_PAGE(!mops
, page
);
103 if (!mops
->isolate_page(page
, mode
))
104 goto out_no_isolated
;
106 /* Driver shouldn't use PG_isolated bit of page->flags */
107 WARN_ON_ONCE(PageIsolated(page
));
108 SetPageIsolated(page
);
121 static void putback_movable_page(struct page
*page
)
123 const struct movable_operations
*mops
= page_movable_ops(page
);
125 mops
->putback_page(page
);
126 ClearPageIsolated(page
);
130 * Put previously isolated pages back onto the appropriate lists
131 * from where they were once taken off for compaction/migration.
133 * This function shall be used whenever the isolated pageset has been
134 * built from lru, balloon, hugetlbfs page. See isolate_migratepages_range()
135 * and isolate_hugetlb().
137 void putback_movable_pages(struct list_head
*l
)
142 list_for_each_entry_safe(page
, page2
, l
, lru
) {
143 if (unlikely(PageHuge(page
))) {
144 putback_active_hugepage(page
);
147 list_del(&page
->lru
);
149 * We isolated non-lru movable page so here we can use
150 * __PageMovable because LRU page's mapping cannot have
151 * PAGE_MAPPING_MOVABLE.
153 if (unlikely(__PageMovable(page
))) {
154 VM_BUG_ON_PAGE(!PageIsolated(page
), page
);
156 if (PageMovable(page
))
157 putback_movable_page(page
);
159 ClearPageIsolated(page
);
163 mod_node_page_state(page_pgdat(page
), NR_ISOLATED_ANON
+
164 page_is_file_lru(page
), -thp_nr_pages(page
));
165 putback_lru_page(page
);
171 * Restore a potential migration pte to a working pte entry
173 static bool remove_migration_pte(struct folio
*folio
,
174 struct vm_area_struct
*vma
, unsigned long addr
, void *old
)
176 DEFINE_FOLIO_VMA_WALK(pvmw
, old
, vma
, addr
, PVMW_SYNC
| PVMW_MIGRATION
);
178 while (page_vma_mapped_walk(&pvmw
)) {
179 rmap_t rmap_flags
= RMAP_NONE
;
183 unsigned long idx
= 0;
185 /* pgoff is invalid for ksm pages, but they are never large */
186 if (folio_test_large(folio
) && !folio_test_hugetlb(folio
))
187 idx
= linear_page_index(vma
, pvmw
.address
) - pvmw
.pgoff
;
188 new = folio_page(folio
, idx
);
190 #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
191 /* PMD-mapped THP migration entry */
193 VM_BUG_ON_FOLIO(folio_test_hugetlb(folio
) ||
194 !folio_test_pmd_mappable(folio
), folio
);
195 remove_migration_pmd(&pvmw
, new);
201 pte
= pte_mkold(mk_pte(new, READ_ONCE(vma
->vm_page_prot
)));
202 if (pte_swp_soft_dirty(*pvmw
.pte
))
203 pte
= pte_mksoft_dirty(pte
);
206 * Recheck VMA as permissions can change since migration started
208 entry
= pte_to_swp_entry(*pvmw
.pte
);
209 if (is_writable_migration_entry(entry
))
210 pte
= maybe_mkwrite(pte
, vma
);
211 else if (pte_swp_uffd_wp(*pvmw
.pte
))
212 pte
= pte_mkuffd_wp(pte
);
214 if (folio_test_anon(folio
) && !is_readable_migration_entry(entry
))
215 rmap_flags
|= RMAP_EXCLUSIVE
;
217 if (unlikely(is_device_private_page(new))) {
219 entry
= make_writable_device_private_entry(
222 entry
= make_readable_device_private_entry(
224 pte
= swp_entry_to_pte(entry
);
225 if (pte_swp_soft_dirty(*pvmw
.pte
))
226 pte
= pte_swp_mksoft_dirty(pte
);
227 if (pte_swp_uffd_wp(*pvmw
.pte
))
228 pte
= pte_swp_mkuffd_wp(pte
);
231 #ifdef CONFIG_HUGETLB_PAGE
232 if (folio_test_hugetlb(folio
)) {
233 unsigned int shift
= huge_page_shift(hstate_vma(vma
));
235 pte
= pte_mkhuge(pte
);
236 pte
= arch_make_huge_pte(pte
, shift
, vma
->vm_flags
);
237 if (folio_test_anon(folio
))
238 hugepage_add_anon_rmap(new, vma
, pvmw
.address
,
241 page_dup_file_rmap(new, true);
242 set_huge_pte_at(vma
->vm_mm
, pvmw
.address
, pvmw
.pte
, pte
);
246 if (folio_test_anon(folio
))
247 page_add_anon_rmap(new, vma
, pvmw
.address
,
250 page_add_file_rmap(new, vma
, false);
251 set_pte_at(vma
->vm_mm
, pvmw
.address
, pvmw
.pte
, pte
);
253 if (vma
->vm_flags
& VM_LOCKED
)
254 mlock_page_drain_local();
256 trace_remove_migration_pte(pvmw
.address
, pte_val(pte
),
257 compound_order(new));
259 /* No need to invalidate - it was non-present before */
260 update_mmu_cache(vma
, pvmw
.address
, pvmw
.pte
);
267 * Get rid of all migration entries and replace them by
268 * references to the indicated page.
270 void remove_migration_ptes(struct folio
*src
, struct folio
*dst
, bool locked
)
272 struct rmap_walk_control rwc
= {
273 .rmap_one
= remove_migration_pte
,
278 rmap_walk_locked(dst
, &rwc
);
280 rmap_walk(dst
, &rwc
);
284 * Something used the pte of a page under migration. We need to
285 * get to the page and wait until migration is finished.
286 * When we return from this function the fault will be retried.
288 void __migration_entry_wait(struct mm_struct
*mm
, pte_t
*ptep
,
296 if (!is_swap_pte(pte
))
299 entry
= pte_to_swp_entry(pte
);
300 if (!is_migration_entry(entry
))
303 migration_entry_wait_on_locked(entry
, ptep
, ptl
);
306 pte_unmap_unlock(ptep
, ptl
);
309 void migration_entry_wait(struct mm_struct
*mm
, pmd_t
*pmd
,
310 unsigned long address
)
312 spinlock_t
*ptl
= pte_lockptr(mm
, pmd
);
313 pte_t
*ptep
= pte_offset_map(pmd
, address
);
314 __migration_entry_wait(mm
, ptep
, ptl
);
317 #ifdef CONFIG_HUGETLB_PAGE
318 void __migration_entry_wait_huge(pte_t
*ptep
, spinlock_t
*ptl
)
323 pte
= huge_ptep_get(ptep
);
325 if (unlikely(!is_hugetlb_entry_migration(pte
)))
328 migration_entry_wait_on_locked(pte_to_swp_entry(pte
), NULL
, ptl
);
331 void migration_entry_wait_huge(struct vm_area_struct
*vma
, pte_t
*pte
)
333 spinlock_t
*ptl
= huge_pte_lockptr(hstate_vma(vma
), vma
->vm_mm
, pte
);
335 __migration_entry_wait_huge(pte
, ptl
);
339 #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
340 void pmd_migration_entry_wait(struct mm_struct
*mm
, pmd_t
*pmd
)
344 ptl
= pmd_lock(mm
, pmd
);
345 if (!is_pmd_migration_entry(*pmd
))
347 migration_entry_wait_on_locked(pmd_to_swp_entry(*pmd
), NULL
, ptl
);
354 static int folio_expected_refs(struct address_space
*mapping
,
361 refs
+= folio_nr_pages(folio
);
362 if (folio_test_private(folio
))
369 * Replace the page in the mapping.
371 * The number of remaining references must be:
372 * 1 for anonymous pages without a mapping
373 * 2 for pages with a mapping
374 * 3 for pages with a mapping and PagePrivate/PagePrivate2 set.
376 int folio_migrate_mapping(struct address_space
*mapping
,
377 struct folio
*newfolio
, struct folio
*folio
, int extra_count
)
379 XA_STATE(xas
, &mapping
->i_pages
, folio_index(folio
));
380 struct zone
*oldzone
, *newzone
;
382 int expected_count
= folio_expected_refs(mapping
, folio
) + extra_count
;
383 long nr
= folio_nr_pages(folio
);
386 /* Anonymous page without mapping */
387 if (folio_ref_count(folio
) != expected_count
)
390 /* No turning back from here */
391 newfolio
->index
= folio
->index
;
392 newfolio
->mapping
= folio
->mapping
;
393 if (folio_test_swapbacked(folio
))
394 __folio_set_swapbacked(newfolio
);
396 return MIGRATEPAGE_SUCCESS
;
399 oldzone
= folio_zone(folio
);
400 newzone
= folio_zone(newfolio
);
403 if (!folio_ref_freeze(folio
, expected_count
)) {
404 xas_unlock_irq(&xas
);
409 * Now we know that no one else is looking at the folio:
410 * no turning back from here.
412 newfolio
->index
= folio
->index
;
413 newfolio
->mapping
= folio
->mapping
;
414 folio_ref_add(newfolio
, nr
); /* add cache reference */
415 if (folio_test_swapbacked(folio
)) {
416 __folio_set_swapbacked(newfolio
);
417 if (folio_test_swapcache(folio
)) {
418 folio_set_swapcache(newfolio
);
419 newfolio
->private = folio_get_private(folio
);
422 VM_BUG_ON_FOLIO(folio_test_swapcache(folio
), folio
);
425 /* Move dirty while page refs frozen and newpage not yet exposed */
426 dirty
= folio_test_dirty(folio
);
428 folio_clear_dirty(folio
);
429 folio_set_dirty(newfolio
);
432 xas_store(&xas
, newfolio
);
435 * Drop cache reference from old page by unfreezing
436 * to one less reference.
437 * We know this isn't the last reference.
439 folio_ref_unfreeze(folio
, expected_count
- nr
);
442 /* Leave irq disabled to prevent preemption while updating stats */
445 * If moved to a different zone then also account
446 * the page for that zone. Other VM counters will be
447 * taken care of when we establish references to the
448 * new page and drop references to the old page.
450 * Note that anonymous pages are accounted for
451 * via NR_FILE_PAGES and NR_ANON_MAPPED if they
452 * are mapped to swap space.
454 if (newzone
!= oldzone
) {
455 struct lruvec
*old_lruvec
, *new_lruvec
;
456 struct mem_cgroup
*memcg
;
458 memcg
= folio_memcg(folio
);
459 old_lruvec
= mem_cgroup_lruvec(memcg
, oldzone
->zone_pgdat
);
460 new_lruvec
= mem_cgroup_lruvec(memcg
, newzone
->zone_pgdat
);
462 __mod_lruvec_state(old_lruvec
, NR_FILE_PAGES
, -nr
);
463 __mod_lruvec_state(new_lruvec
, NR_FILE_PAGES
, nr
);
464 if (folio_test_swapbacked(folio
) && !folio_test_swapcache(folio
)) {
465 __mod_lruvec_state(old_lruvec
, NR_SHMEM
, -nr
);
466 __mod_lruvec_state(new_lruvec
, NR_SHMEM
, nr
);
469 if (folio_test_swapcache(folio
)) {
470 __mod_lruvec_state(old_lruvec
, NR_SWAPCACHE
, -nr
);
471 __mod_lruvec_state(new_lruvec
, NR_SWAPCACHE
, nr
);
474 if (dirty
&& mapping_can_writeback(mapping
)) {
475 __mod_lruvec_state(old_lruvec
, NR_FILE_DIRTY
, -nr
);
476 __mod_zone_page_state(oldzone
, NR_ZONE_WRITE_PENDING
, -nr
);
477 __mod_lruvec_state(new_lruvec
, NR_FILE_DIRTY
, nr
);
478 __mod_zone_page_state(newzone
, NR_ZONE_WRITE_PENDING
, nr
);
483 return MIGRATEPAGE_SUCCESS
;
485 EXPORT_SYMBOL(folio_migrate_mapping
);
488 * The expected number of remaining references is the same as that
489 * of folio_migrate_mapping().
491 int migrate_huge_page_move_mapping(struct address_space
*mapping
,
492 struct folio
*dst
, struct folio
*src
)
494 XA_STATE(xas
, &mapping
->i_pages
, folio_index(src
));
498 expected_count
= 2 + folio_has_private(src
);
499 if (!folio_ref_freeze(src
, expected_count
)) {
500 xas_unlock_irq(&xas
);
504 dst
->index
= src
->index
;
505 dst
->mapping
= src
->mapping
;
509 xas_store(&xas
, dst
);
511 folio_ref_unfreeze(src
, expected_count
- 1);
513 xas_unlock_irq(&xas
);
515 return MIGRATEPAGE_SUCCESS
;
519 * Copy the flags and some other ancillary information
521 void folio_migrate_flags(struct folio
*newfolio
, struct folio
*folio
)
525 if (folio_test_error(folio
))
526 folio_set_error(newfolio
);
527 if (folio_test_referenced(folio
))
528 folio_set_referenced(newfolio
);
529 if (folio_test_uptodate(folio
))
530 folio_mark_uptodate(newfolio
);
531 if (folio_test_clear_active(folio
)) {
532 VM_BUG_ON_FOLIO(folio_test_unevictable(folio
), folio
);
533 folio_set_active(newfolio
);
534 } else if (folio_test_clear_unevictable(folio
))
535 folio_set_unevictable(newfolio
);
536 if (folio_test_workingset(folio
))
537 folio_set_workingset(newfolio
);
538 if (folio_test_checked(folio
))
539 folio_set_checked(newfolio
);
541 * PG_anon_exclusive (-> PG_mappedtodisk) is always migrated via
542 * migration entries. We can still have PG_anon_exclusive set on an
543 * effectively unmapped and unreferenced first sub-pages of an
544 * anonymous THP: we can simply copy it here via PG_mappedtodisk.
546 if (folio_test_mappedtodisk(folio
))
547 folio_set_mappedtodisk(newfolio
);
549 /* Move dirty on pages not done by folio_migrate_mapping() */
550 if (folio_test_dirty(folio
))
551 folio_set_dirty(newfolio
);
553 if (folio_test_young(folio
))
554 folio_set_young(newfolio
);
555 if (folio_test_idle(folio
))
556 folio_set_idle(newfolio
);
559 * Copy NUMA information to the new page, to prevent over-eager
560 * future migrations of this same page.
562 cpupid
= page_cpupid_xchg_last(&folio
->page
, -1);
563 page_cpupid_xchg_last(&newfolio
->page
, cpupid
);
565 folio_migrate_ksm(newfolio
, folio
);
567 * Please do not reorder this without considering how mm/ksm.c's
568 * get_ksm_page() depends upon ksm_migrate_page() and PageSwapCache().
570 if (folio_test_swapcache(folio
))
571 folio_clear_swapcache(folio
);
572 folio_clear_private(folio
);
574 /* page->private contains hugetlb specific flags */
575 if (!folio_test_hugetlb(folio
))
576 folio
->private = NULL
;
579 * If any waiters have accumulated on the new page then
582 if (folio_test_writeback(newfolio
))
583 folio_end_writeback(newfolio
);
586 * PG_readahead shares the same bit with PG_reclaim. The above
587 * end_page_writeback() may clear PG_readahead mistakenly, so set the
590 if (folio_test_readahead(folio
))
591 folio_set_readahead(newfolio
);
593 folio_copy_owner(newfolio
, folio
);
595 if (!folio_test_hugetlb(folio
))
596 mem_cgroup_migrate(folio
, newfolio
);
598 EXPORT_SYMBOL(folio_migrate_flags
);
600 void folio_migrate_copy(struct folio
*newfolio
, struct folio
*folio
)
602 folio_copy(newfolio
, folio
);
603 folio_migrate_flags(newfolio
, folio
);
605 EXPORT_SYMBOL(folio_migrate_copy
);
607 /************************************************************
608 * Migration functions
609 ***********************************************************/
612 * migrate_folio() - Simple folio migration.
613 * @mapping: The address_space containing the folio.
614 * @dst: The folio to migrate the data to.
615 * @src: The folio containing the current data.
616 * @mode: How to migrate the page.
618 * Common logic to directly migrate a single LRU folio suitable for
619 * folios that do not use PagePrivate/PagePrivate2.
621 * Folios are locked upon entry and exit.
623 int migrate_folio(struct address_space
*mapping
, struct folio
*dst
,
624 struct folio
*src
, enum migrate_mode mode
)
628 BUG_ON(folio_test_writeback(src
)); /* Writeback must be complete */
630 rc
= folio_migrate_mapping(mapping
, dst
, src
, 0);
632 if (rc
!= MIGRATEPAGE_SUCCESS
)
635 if (mode
!= MIGRATE_SYNC_NO_COPY
)
636 folio_migrate_copy(dst
, src
);
638 folio_migrate_flags(dst
, src
);
639 return MIGRATEPAGE_SUCCESS
;
641 EXPORT_SYMBOL(migrate_folio
);
644 /* Returns true if all buffers are successfully locked */
645 static bool buffer_migrate_lock_buffers(struct buffer_head
*head
,
646 enum migrate_mode mode
)
648 struct buffer_head
*bh
= head
;
650 /* Simple case, sync compaction */
651 if (mode
!= MIGRATE_ASYNC
) {
654 bh
= bh
->b_this_page
;
656 } while (bh
!= head
);
661 /* async case, we cannot block on lock_buffer so use trylock_buffer */
663 if (!trylock_buffer(bh
)) {
665 * We failed to lock the buffer and cannot stall in
666 * async migration. Release the taken locks
668 struct buffer_head
*failed_bh
= bh
;
670 while (bh
!= failed_bh
) {
672 bh
= bh
->b_this_page
;
677 bh
= bh
->b_this_page
;
678 } while (bh
!= head
);
682 static int __buffer_migrate_folio(struct address_space
*mapping
,
683 struct folio
*dst
, struct folio
*src
, enum migrate_mode mode
,
686 struct buffer_head
*bh
, *head
;
690 head
= folio_buffers(src
);
692 return migrate_folio(mapping
, dst
, src
, mode
);
694 /* Check whether page does not have extra refs before we do more work */
695 expected_count
= folio_expected_refs(mapping
, src
);
696 if (folio_ref_count(src
) != expected_count
)
699 if (!buffer_migrate_lock_buffers(head
, mode
))
704 bool invalidated
= false;
708 spin_lock(&mapping
->private_lock
);
711 if (atomic_read(&bh
->b_count
)) {
715 bh
= bh
->b_this_page
;
716 } while (bh
!= head
);
722 spin_unlock(&mapping
->private_lock
);
723 invalidate_bh_lrus();
725 goto recheck_buffers
;
729 rc
= folio_migrate_mapping(mapping
, dst
, src
, 0);
730 if (rc
!= MIGRATEPAGE_SUCCESS
)
733 folio_attach_private(dst
, folio_detach_private(src
));
737 set_bh_page(bh
, &dst
->page
, bh_offset(bh
));
738 bh
= bh
->b_this_page
;
739 } while (bh
!= head
);
741 if (mode
!= MIGRATE_SYNC_NO_COPY
)
742 folio_migrate_copy(dst
, src
);
744 folio_migrate_flags(dst
, src
);
746 rc
= MIGRATEPAGE_SUCCESS
;
749 spin_unlock(&mapping
->private_lock
);
753 bh
= bh
->b_this_page
;
754 } while (bh
!= head
);
760 * buffer_migrate_folio() - Migration function for folios with buffers.
761 * @mapping: The address space containing @src.
762 * @dst: The folio to migrate to.
763 * @src: The folio to migrate from.
764 * @mode: How to migrate the folio.
766 * This function can only be used if the underlying filesystem guarantees
767 * that no other references to @src exist. For example attached buffer
768 * heads are accessed only under the folio lock. If your filesystem cannot
769 * provide this guarantee, buffer_migrate_folio_norefs() may be more
772 * Return: 0 on success or a negative errno on failure.
774 int buffer_migrate_folio(struct address_space
*mapping
,
775 struct folio
*dst
, struct folio
*src
, enum migrate_mode mode
)
777 return __buffer_migrate_folio(mapping
, dst
, src
, mode
, false);
779 EXPORT_SYMBOL(buffer_migrate_folio
);
782 * buffer_migrate_folio_norefs() - Migration function for folios with buffers.
783 * @mapping: The address space containing @src.
784 * @dst: The folio to migrate to.
785 * @src: The folio to migrate from.
786 * @mode: How to migrate the folio.
788 * Like buffer_migrate_folio() except that this variant is more careful
789 * and checks that there are also no buffer head references. This function
790 * is the right one for mappings where buffer heads are directly looked
791 * up and referenced (such as block device mappings).
793 * Return: 0 on success or a negative errno on failure.
795 int buffer_migrate_folio_norefs(struct address_space
*mapping
,
796 struct folio
*dst
, struct folio
*src
, enum migrate_mode mode
)
798 return __buffer_migrate_folio(mapping
, dst
, src
, mode
, true);
802 int filemap_migrate_folio(struct address_space
*mapping
,
803 struct folio
*dst
, struct folio
*src
, enum migrate_mode mode
)
807 ret
= folio_migrate_mapping(mapping
, dst
, src
, 0);
808 if (ret
!= MIGRATEPAGE_SUCCESS
)
811 if (folio_get_private(src
))
812 folio_attach_private(dst
, folio_detach_private(src
));
814 if (mode
!= MIGRATE_SYNC_NO_COPY
)
815 folio_migrate_copy(dst
, src
);
817 folio_migrate_flags(dst
, src
);
818 return MIGRATEPAGE_SUCCESS
;
820 EXPORT_SYMBOL_GPL(filemap_migrate_folio
);
823 * Writeback a folio to clean the dirty state
825 static int writeout(struct address_space
*mapping
, struct folio
*folio
)
827 struct writeback_control wbc
= {
828 .sync_mode
= WB_SYNC_NONE
,
831 .range_end
= LLONG_MAX
,
836 if (!mapping
->a_ops
->writepage
)
837 /* No write method for the address space */
840 if (!folio_clear_dirty_for_io(folio
))
841 /* Someone else already triggered a write */
845 * A dirty folio may imply that the underlying filesystem has
846 * the folio on some queue. So the folio must be clean for
847 * migration. Writeout may mean we lose the lock and the
848 * folio state is no longer what we checked for earlier.
849 * At this point we know that the migration attempt cannot
852 remove_migration_ptes(folio
, folio
, false);
854 rc
= mapping
->a_ops
->writepage(&folio
->page
, &wbc
);
856 if (rc
!= AOP_WRITEPAGE_ACTIVATE
)
857 /* unlocked. Relock */
860 return (rc
< 0) ? -EIO
: -EAGAIN
;
864 * Default handling if a filesystem does not provide a migration function.
866 static int fallback_migrate_folio(struct address_space
*mapping
,
867 struct folio
*dst
, struct folio
*src
, enum migrate_mode mode
)
869 if (folio_test_dirty(src
)) {
870 /* Only writeback folios in full synchronous migration */
873 case MIGRATE_SYNC_NO_COPY
:
878 return writeout(mapping
, src
);
882 * Buffers may be managed in a filesystem specific way.
883 * We must have no buffers or drop them.
885 if (folio_test_private(src
) &&
886 !filemap_release_folio(src
, GFP_KERNEL
))
887 return mode
== MIGRATE_SYNC
? -EAGAIN
: -EBUSY
;
889 return migrate_folio(mapping
, dst
, src
, mode
);
893 * Move a page to a newly allocated page
894 * The page is locked and all ptes have been successfully removed.
896 * The new page will have replaced the old page if this function
901 * MIGRATEPAGE_SUCCESS - success
903 static int move_to_new_folio(struct folio
*dst
, struct folio
*src
,
904 enum migrate_mode mode
)
907 bool is_lru
= !__PageMovable(&src
->page
);
909 VM_BUG_ON_FOLIO(!folio_test_locked(src
), src
);
910 VM_BUG_ON_FOLIO(!folio_test_locked(dst
), dst
);
912 if (likely(is_lru
)) {
913 struct address_space
*mapping
= folio_mapping(src
);
916 rc
= migrate_folio(mapping
, dst
, src
, mode
);
917 else if (mapping
->a_ops
->migrate_folio
)
919 * Most folios have a mapping and most filesystems
920 * provide a migrate_folio callback. Anonymous folios
921 * are part of swap space which also has its own
922 * migrate_folio callback. This is the most common path
923 * for page migration.
925 rc
= mapping
->a_ops
->migrate_folio(mapping
, dst
, src
,
928 rc
= fallback_migrate_folio(mapping
, dst
, src
, mode
);
930 const struct movable_operations
*mops
;
933 * In case of non-lru page, it could be released after
934 * isolation step. In that case, we shouldn't try migration.
936 VM_BUG_ON_FOLIO(!folio_test_isolated(src
), src
);
937 if (!folio_test_movable(src
)) {
938 rc
= MIGRATEPAGE_SUCCESS
;
939 folio_clear_isolated(src
);
943 mops
= page_movable_ops(&src
->page
);
944 rc
= mops
->migrate_page(&dst
->page
, &src
->page
, mode
);
945 WARN_ON_ONCE(rc
== MIGRATEPAGE_SUCCESS
&&
946 !folio_test_isolated(src
));
950 * When successful, old pagecache src->mapping must be cleared before
951 * src is freed; but stats require that PageAnon be left as PageAnon.
953 if (rc
== MIGRATEPAGE_SUCCESS
) {
954 if (__PageMovable(&src
->page
)) {
955 VM_BUG_ON_FOLIO(!folio_test_isolated(src
), src
);
958 * We clear PG_movable under page_lock so any compactor
959 * cannot try to migrate this page.
961 folio_clear_isolated(src
);
965 * Anonymous and movable src->mapping will be cleared by
966 * free_pages_prepare so don't reset it here for keeping
967 * the type to work PageAnon, for example.
969 if (!folio_mapping_flags(src
))
972 if (likely(!folio_is_zone_device(dst
)))
973 flush_dcache_folio(dst
);
979 static int __unmap_and_move(struct page
*page
, struct page
*newpage
,
980 int force
, enum migrate_mode mode
)
982 struct folio
*folio
= page_folio(page
);
983 struct folio
*dst
= page_folio(newpage
);
985 bool page_was_mapped
= false;
986 struct anon_vma
*anon_vma
= NULL
;
987 bool is_lru
= !__PageMovable(page
);
989 if (!trylock_page(page
)) {
990 if (!force
|| mode
== MIGRATE_ASYNC
)
994 * It's not safe for direct compaction to call lock_page.
995 * For example, during page readahead pages are added locked
996 * to the LRU. Later, when the IO completes the pages are
997 * marked uptodate and unlocked. However, the queueing
998 * could be merging multiple pages for one bio (e.g.
999 * mpage_readahead). If an allocation happens for the
1000 * second or third page, the process can end up locking
1001 * the same page twice and deadlocking. Rather than
1002 * trying to be clever about what pages can be locked,
1003 * avoid the use of lock_page for direct compaction
1006 if (current
->flags
& PF_MEMALLOC
)
1012 if (PageWriteback(page
)) {
1014 * Only in the case of a full synchronous migration is it
1015 * necessary to wait for PageWriteback. In the async case,
1016 * the retry loop is too short and in the sync-light case,
1017 * the overhead of stalling is too much
1021 case MIGRATE_SYNC_NO_COPY
:
1029 wait_on_page_writeback(page
);
1033 * By try_to_migrate(), page->mapcount goes down to 0 here. In this case,
1034 * we cannot notice that anon_vma is freed while we migrates a page.
1035 * This get_anon_vma() delays freeing anon_vma pointer until the end
1036 * of migration. File cache pages are no problem because of page_lock()
1037 * File Caches may use write_page() or lock_page() in migration, then,
1038 * just care Anon page here.
1040 * Only page_get_anon_vma() understands the subtleties of
1041 * getting a hold on an anon_vma from outside one of its mms.
1042 * But if we cannot get anon_vma, then we won't need it anyway,
1043 * because that implies that the anon page is no longer mapped
1044 * (and cannot be remapped so long as we hold the page lock).
1046 if (PageAnon(page
) && !PageKsm(page
))
1047 anon_vma
= page_get_anon_vma(page
);
1050 * Block others from accessing the new page when we get around to
1051 * establishing additional references. We are usually the only one
1052 * holding a reference to newpage at this point. We used to have a BUG
1053 * here if trylock_page(newpage) fails, but would like to allow for
1054 * cases where there might be a race with the previous use of newpage.
1055 * This is much like races on refcount of oldpage: just don't BUG().
1057 if (unlikely(!trylock_page(newpage
)))
1060 if (unlikely(!is_lru
)) {
1061 rc
= move_to_new_folio(dst
, folio
, mode
);
1062 goto out_unlock_both
;
1066 * Corner case handling:
1067 * 1. When a new swap-cache page is read into, it is added to the LRU
1068 * and treated as swapcache but it has no rmap yet.
1069 * Calling try_to_unmap() against a page->mapping==NULL page will
1070 * trigger a BUG. So handle it here.
1071 * 2. An orphaned page (see truncate_cleanup_page) might have
1072 * fs-private metadata. The page can be picked up due to memory
1073 * offlining. Everywhere else except page reclaim, the page is
1074 * invisible to the vm, so the page can not be migrated. So try to
1075 * free the metadata, so the page can be freed.
1077 if (!page
->mapping
) {
1078 VM_BUG_ON_PAGE(PageAnon(page
), page
);
1079 if (page_has_private(page
)) {
1080 try_to_free_buffers(folio
);
1081 goto out_unlock_both
;
1083 } else if (page_mapped(page
)) {
1084 /* Establish migration ptes */
1085 VM_BUG_ON_PAGE(PageAnon(page
) && !PageKsm(page
) && !anon_vma
,
1087 try_to_migrate(folio
, 0);
1088 page_was_mapped
= true;
1091 if (!page_mapped(page
))
1092 rc
= move_to_new_folio(dst
, folio
, mode
);
1095 * When successful, push newpage to LRU immediately: so that if it
1096 * turns out to be an mlocked page, remove_migration_ptes() will
1097 * automatically build up the correct newpage->mlock_count for it.
1099 * We would like to do something similar for the old page, when
1100 * unsuccessful, and other cases when a page has been temporarily
1101 * isolated from the unevictable LRU: but this case is the easiest.
1103 if (rc
== MIGRATEPAGE_SUCCESS
) {
1104 lru_cache_add(newpage
);
1105 if (page_was_mapped
)
1109 if (page_was_mapped
)
1110 remove_migration_ptes(folio
,
1111 rc
== MIGRATEPAGE_SUCCESS
? dst
: folio
, false);
1114 unlock_page(newpage
);
1116 /* Drop an anon_vma reference if we took one */
1118 put_anon_vma(anon_vma
);
1122 * If migration is successful, decrease refcount of the newpage,
1123 * which will not free the page because new page owner increased
1126 if (rc
== MIGRATEPAGE_SUCCESS
)
1133 * Obtain the lock on page, remove all ptes and migrate the page
1134 * to the newly allocated page in newpage.
1136 static int unmap_and_move(new_page_t get_new_page
,
1137 free_page_t put_new_page
,
1138 unsigned long private, struct page
*page
,
1139 int force
, enum migrate_mode mode
,
1140 enum migrate_reason reason
,
1141 struct list_head
*ret
)
1143 int rc
= MIGRATEPAGE_SUCCESS
;
1144 struct page
*newpage
= NULL
;
1146 if (!thp_migration_supported() && PageTransHuge(page
))
1149 if (page_count(page
) == 1) {
1150 /* Page was freed from under us. So we are done. */
1151 ClearPageActive(page
);
1152 ClearPageUnevictable(page
);
1153 /* free_pages_prepare() will clear PG_isolated. */
1157 newpage
= get_new_page(page
, private);
1161 newpage
->private = 0;
1162 rc
= __unmap_and_move(page
, newpage
, force
, mode
);
1163 if (rc
== MIGRATEPAGE_SUCCESS
)
1164 set_page_owner_migrate_reason(newpage
, reason
);
1167 if (rc
!= -EAGAIN
) {
1169 * A page that has been migrated has all references
1170 * removed and will be freed. A page that has not been
1171 * migrated will have kept its references and be restored.
1173 list_del(&page
->lru
);
1177 * If migration is successful, releases reference grabbed during
1178 * isolation. Otherwise, restore the page to right list unless
1181 if (rc
== MIGRATEPAGE_SUCCESS
) {
1183 * Compaction can migrate also non-LRU pages which are
1184 * not accounted to NR_ISOLATED_*. They can be recognized
1187 if (likely(!__PageMovable(page
)))
1188 mod_node_page_state(page_pgdat(page
), NR_ISOLATED_ANON
+
1189 page_is_file_lru(page
), -thp_nr_pages(page
));
1191 if (reason
!= MR_MEMORY_FAILURE
)
1193 * We release the page in page_handle_poison.
1198 list_add_tail(&page
->lru
, ret
);
1201 put_new_page(newpage
, private);
1210 * Counterpart of unmap_and_move_page() for hugepage migration.
1212 * This function doesn't wait the completion of hugepage I/O
1213 * because there is no race between I/O and migration for hugepage.
1214 * Note that currently hugepage I/O occurs only in direct I/O
1215 * where no lock is held and PG_writeback is irrelevant,
1216 * and writeback status of all subpages are counted in the reference
1217 * count of the head page (i.e. if all subpages of a 2MB hugepage are
1218 * under direct I/O, the reference of the head page is 512 and a bit more.)
1219 * This means that when we try to migrate hugepage whose subpages are
1220 * doing direct I/O, some references remain after try_to_unmap() and
1221 * hugepage migration fails without data corruption.
1223 * There is also no race when direct I/O is issued on the page under migration,
1224 * because then pte is replaced with migration swap entry and direct I/O code
1225 * will wait in the page fault for migration to complete.
1227 static int unmap_and_move_huge_page(new_page_t get_new_page
,
1228 free_page_t put_new_page
, unsigned long private,
1229 struct page
*hpage
, int force
,
1230 enum migrate_mode mode
, int reason
,
1231 struct list_head
*ret
)
1233 struct folio
*dst
, *src
= page_folio(hpage
);
1235 int page_was_mapped
= 0;
1236 struct page
*new_hpage
;
1237 struct anon_vma
*anon_vma
= NULL
;
1238 struct address_space
*mapping
= NULL
;
1241 * Migratability of hugepages depends on architectures and their size.
1242 * This check is necessary because some callers of hugepage migration
1243 * like soft offline and memory hotremove don't walk through page
1244 * tables or check whether the hugepage is pmd-based or not before
1245 * kicking migration.
1247 if (!hugepage_migration_supported(page_hstate(hpage
))) {
1248 list_move_tail(&hpage
->lru
, ret
);
1252 if (page_count(hpage
) == 1) {
1253 /* page was freed from under us. So we are done. */
1254 putback_active_hugepage(hpage
);
1255 return MIGRATEPAGE_SUCCESS
;
1258 new_hpage
= get_new_page(hpage
, private);
1261 dst
= page_folio(new_hpage
);
1263 if (!trylock_page(hpage
)) {
1268 case MIGRATE_SYNC_NO_COPY
:
1277 * Check for pages which are in the process of being freed. Without
1278 * page_mapping() set, hugetlbfs specific move page routine will not
1279 * be called and we could leak usage counts for subpools.
1281 if (hugetlb_page_subpool(hpage
) && !page_mapping(hpage
)) {
1286 if (PageAnon(hpage
))
1287 anon_vma
= page_get_anon_vma(hpage
);
1289 if (unlikely(!trylock_page(new_hpage
)))
1292 if (page_mapped(hpage
)) {
1293 enum ttu_flags ttu
= 0;
1295 if (!PageAnon(hpage
)) {
1297 * In shared mappings, try_to_unmap could potentially
1298 * call huge_pmd_unshare. Because of this, take
1299 * semaphore in write mode here and set TTU_RMAP_LOCKED
1300 * to let lower levels know we have taken the lock.
1302 mapping
= hugetlb_page_mapping_lock_write(hpage
);
1303 if (unlikely(!mapping
))
1304 goto unlock_put_anon
;
1306 ttu
= TTU_RMAP_LOCKED
;
1309 try_to_migrate(src
, ttu
);
1310 page_was_mapped
= 1;
1312 if (ttu
& TTU_RMAP_LOCKED
)
1313 i_mmap_unlock_write(mapping
);
1316 if (!page_mapped(hpage
))
1317 rc
= move_to_new_folio(dst
, src
, mode
);
1319 if (page_was_mapped
)
1320 remove_migration_ptes(src
,
1321 rc
== MIGRATEPAGE_SUCCESS
? dst
: src
, false);
1324 unlock_page(new_hpage
);
1328 put_anon_vma(anon_vma
);
1330 if (rc
== MIGRATEPAGE_SUCCESS
) {
1331 move_hugetlb_state(hpage
, new_hpage
, reason
);
1332 put_new_page
= NULL
;
1338 if (rc
== MIGRATEPAGE_SUCCESS
)
1339 putback_active_hugepage(hpage
);
1340 else if (rc
!= -EAGAIN
)
1341 list_move_tail(&hpage
->lru
, ret
);
1344 * If migration was not successful and there's a freeing callback, use
1345 * it. Otherwise, put_page() will drop the reference grabbed during
1349 put_new_page(new_hpage
, private);
1351 putback_active_hugepage(new_hpage
);
1356 static inline int try_split_thp(struct page
*page
, struct page
**page2
,
1357 struct list_head
*from
)
1362 rc
= split_huge_page_to_list(page
, from
);
1365 list_safe_reset_next(page
, *page2
, lru
);
1371 * migrate_pages - migrate the pages specified in a list, to the free pages
1372 * supplied as the target for the page migration
1374 * @from: The list of pages to be migrated.
1375 * @get_new_page: The function used to allocate free pages to be used
1376 * as the target of the page migration.
1377 * @put_new_page: The function used to free target pages if migration
1378 * fails, or NULL if no special handling is necessary.
1379 * @private: Private data to be passed on to get_new_page()
1380 * @mode: The migration mode that specifies the constraints for
1381 * page migration, if any.
1382 * @reason: The reason for page migration.
1383 * @ret_succeeded: Set to the number of normal pages migrated successfully if
1384 * the caller passes a non-NULL pointer.
1386 * The function returns after 10 attempts or if no pages are movable any more
1387 * because the list has become empty or no retryable pages exist any more.
1388 * It is caller's responsibility to call putback_movable_pages() to return pages
1389 * to the LRU or free list only if ret != 0.
1391 * Returns the number of {normal page, THP, hugetlb} that were not migrated, or
1392 * an error code. The number of THP splits will be considered as the number of
1393 * non-migrated THP, no matter how many subpages of the THP are migrated successfully.
1395 int migrate_pages(struct list_head
*from
, new_page_t get_new_page
,
1396 free_page_t put_new_page
, unsigned long private,
1397 enum migrate_mode mode
, int reason
, unsigned int *ret_succeeded
)
1402 int nr_failed_pages
= 0;
1403 int nr_succeeded
= 0;
1404 int nr_thp_succeeded
= 0;
1405 int nr_thp_failed
= 0;
1406 int nr_thp_split
= 0;
1408 bool is_thp
= false;
1411 int rc
, nr_subpages
;
1412 LIST_HEAD(ret_pages
);
1413 LIST_HEAD(thp_split_pages
);
1414 bool nosplit
= (reason
== MR_NUMA_MISPLACED
);
1415 bool no_subpage_counting
= false;
1417 trace_mm_migrate_pages_start(mode
, reason
);
1419 thp_subpage_migration
:
1420 for (pass
= 0; pass
< 10 && (retry
|| thp_retry
); pass
++) {
1424 list_for_each_entry_safe(page
, page2
, from
, lru
) {
1427 * THP statistics is based on the source huge page.
1428 * Capture required information that might get lost
1431 is_thp
= PageTransHuge(page
) && !PageHuge(page
);
1432 nr_subpages
= compound_nr(page
);
1436 rc
= unmap_and_move_huge_page(get_new_page
,
1437 put_new_page
, private, page
,
1438 pass
> 2, mode
, reason
,
1441 rc
= unmap_and_move(get_new_page
, put_new_page
,
1442 private, page
, pass
> 2, mode
,
1443 reason
, &ret_pages
);
1446 * Success: non hugetlb page will be freed, hugetlb
1447 * page will be put back
1448 * -EAGAIN: stay on the from list
1449 * -ENOMEM: stay on the from list
1450 * Other errno: put on ret_pages list then splice to
1455 * THP migration might be unsupported or the
1456 * allocation could've failed so we should
1457 * retry on the same page with the THP split
1460 * Head page is retried immediately and tail
1461 * pages are added to the tail of the list so
1462 * we encounter them after the rest of the list
1466 /* THP migration is unsupported */
1469 if (!try_split_thp(page
, &page2
, &thp_split_pages
)) {
1473 /* Hugetlb migration is unsupported */
1474 } else if (!no_subpage_counting
) {
1478 nr_failed_pages
+= nr_subpages
;
1482 * When memory is low, don't bother to try to migrate
1483 * other pages, just exit.
1484 * THP NUMA faulting doesn't split THP to retry.
1486 if (is_thp
&& !nosplit
) {
1488 if (!try_split_thp(page
, &page2
, &thp_split_pages
)) {
1492 } else if (!no_subpage_counting
) {
1496 nr_failed_pages
+= nr_subpages
;
1498 * There might be some subpages of fail-to-migrate THPs
1499 * left in thp_split_pages list. Move them back to migration
1500 * list so that they could be put back to the right list by
1501 * the caller otherwise the page refcnt will be leaked.
1503 list_splice_init(&thp_split_pages
, from
);
1504 nr_thp_failed
+= thp_retry
;
1512 case MIGRATEPAGE_SUCCESS
:
1513 nr_succeeded
+= nr_subpages
;
1519 * Permanent failure (-EBUSY, etc.):
1520 * unlike -EAGAIN case, the failed page is
1521 * removed from migration page list and not
1522 * retried in the next outer loop.
1526 else if (!no_subpage_counting
)
1529 nr_failed_pages
+= nr_subpages
;
1535 nr_thp_failed
+= thp_retry
;
1537 * Try to migrate subpages of fail-to-migrate THPs, no nr_failed
1538 * counting in this round, since all subpages of a THP is counted
1539 * as 1 failure in the first round.
1541 if (!list_empty(&thp_split_pages
)) {
1543 * Move non-migrated pages (after 10 retries) to ret_pages
1544 * to avoid migrating them again.
1546 list_splice_init(from
, &ret_pages
);
1547 list_splice_init(&thp_split_pages
, from
);
1548 no_subpage_counting
= true;
1550 goto thp_subpage_migration
;
1553 rc
= nr_failed
+ nr_thp_failed
;
1556 * Put the permanent failure page back to migration list, they
1557 * will be put back to the right list by the caller.
1559 list_splice(&ret_pages
, from
);
1561 count_vm_events(PGMIGRATE_SUCCESS
, nr_succeeded
);
1562 count_vm_events(PGMIGRATE_FAIL
, nr_failed_pages
);
1563 count_vm_events(THP_MIGRATION_SUCCESS
, nr_thp_succeeded
);
1564 count_vm_events(THP_MIGRATION_FAIL
, nr_thp_failed
);
1565 count_vm_events(THP_MIGRATION_SPLIT
, nr_thp_split
);
1566 trace_mm_migrate_pages(nr_succeeded
, nr_failed_pages
, nr_thp_succeeded
,
1567 nr_thp_failed
, nr_thp_split
, mode
, reason
);
1570 *ret_succeeded
= nr_succeeded
;
1575 struct page
*alloc_migration_target(struct page
*page
, unsigned long private)
1577 struct folio
*folio
= page_folio(page
);
1578 struct migration_target_control
*mtc
;
1580 unsigned int order
= 0;
1581 struct folio
*new_folio
= NULL
;
1585 mtc
= (struct migration_target_control
*)private;
1586 gfp_mask
= mtc
->gfp_mask
;
1588 if (nid
== NUMA_NO_NODE
)
1589 nid
= folio_nid(folio
);
1591 if (folio_test_hugetlb(folio
)) {
1592 struct hstate
*h
= page_hstate(&folio
->page
);
1594 gfp_mask
= htlb_modify_alloc_mask(h
, gfp_mask
);
1595 return alloc_huge_page_nodemask(h
, nid
, mtc
->nmask
, gfp_mask
);
1598 if (folio_test_large(folio
)) {
1600 * clear __GFP_RECLAIM to make the migration callback
1601 * consistent with regular THP allocations.
1603 gfp_mask
&= ~__GFP_RECLAIM
;
1604 gfp_mask
|= GFP_TRANSHUGE
;
1605 order
= folio_order(folio
);
1607 zidx
= zone_idx(folio_zone(folio
));
1608 if (is_highmem_idx(zidx
) || zidx
== ZONE_MOVABLE
)
1609 gfp_mask
|= __GFP_HIGHMEM
;
1611 new_folio
= __folio_alloc(gfp_mask
, order
, nid
, mtc
->nmask
);
1613 return &new_folio
->page
;
1618 static int store_status(int __user
*status
, int start
, int value
, int nr
)
1621 if (put_user(value
, status
+ start
))
1629 static int do_move_pages_to_node(struct mm_struct
*mm
,
1630 struct list_head
*pagelist
, int node
)
1633 struct migration_target_control mtc
= {
1635 .gfp_mask
= GFP_HIGHUSER_MOVABLE
| __GFP_THISNODE
,
1638 err
= migrate_pages(pagelist
, alloc_migration_target
, NULL
,
1639 (unsigned long)&mtc
, MIGRATE_SYNC
, MR_SYSCALL
, NULL
);
1641 putback_movable_pages(pagelist
);
1646 * Resolves the given address to a struct page, isolates it from the LRU and
1647 * puts it to the given pagelist.
1649 * errno - if the page cannot be found/isolated
1650 * 0 - when it doesn't have to be migrated because it is already on the
1652 * 1 - when it has been queued
1654 static int add_page_for_migration(struct mm_struct
*mm
, unsigned long addr
,
1655 int node
, struct list_head
*pagelist
, bool migrate_all
)
1657 struct vm_area_struct
*vma
;
1663 vma
= vma_lookup(mm
, addr
);
1664 if (!vma
|| !vma_migratable(vma
))
1667 /* FOLL_DUMP to ignore special (like zero) pages */
1668 page
= follow_page(vma
, addr
, FOLL_GET
| FOLL_DUMP
);
1670 err
= PTR_ERR(page
);
1675 if (!page
|| is_zone_device_page(page
))
1679 if (page_to_nid(page
) == node
)
1683 if (page_mapcount(page
) > 1 && !migrate_all
)
1686 if (PageHuge(page
)) {
1687 if (PageHead(page
)) {
1688 err
= isolate_hugetlb(page
, pagelist
);
1695 head
= compound_head(page
);
1696 err
= isolate_lru_page(head
);
1701 list_add_tail(&head
->lru
, pagelist
);
1702 mod_node_page_state(page_pgdat(head
),
1703 NR_ISOLATED_ANON
+ page_is_file_lru(head
),
1704 thp_nr_pages(head
));
1708 * Either remove the duplicate refcount from
1709 * isolate_lru_page() or drop the page ref if it was
1714 mmap_read_unlock(mm
);
1718 static int move_pages_and_store_status(struct mm_struct
*mm
, int node
,
1719 struct list_head
*pagelist
, int __user
*status
,
1720 int start
, int i
, unsigned long nr_pages
)
1724 if (list_empty(pagelist
))
1727 err
= do_move_pages_to_node(mm
, pagelist
, node
);
1730 * Positive err means the number of failed
1731 * pages to migrate. Since we are going to
1732 * abort and return the number of non-migrated
1733 * pages, so need to include the rest of the
1734 * nr_pages that have not been attempted as
1738 err
+= nr_pages
- i
- 1;
1741 return store_status(status
, start
, node
, i
- start
);
1745 * Migrate an array of page address onto an array of nodes and fill
1746 * the corresponding array of status.
1748 static int do_pages_move(struct mm_struct
*mm
, nodemask_t task_nodes
,
1749 unsigned long nr_pages
,
1750 const void __user
* __user
*pages
,
1751 const int __user
*nodes
,
1752 int __user
*status
, int flags
)
1754 int current_node
= NUMA_NO_NODE
;
1755 LIST_HEAD(pagelist
);
1759 lru_cache_disable();
1761 for (i
= start
= 0; i
< nr_pages
; i
++) {
1762 const void __user
*p
;
1767 if (get_user(p
, pages
+ i
))
1769 if (get_user(node
, nodes
+ i
))
1771 addr
= (unsigned long)untagged_addr(p
);
1774 if (node
< 0 || node
>= MAX_NUMNODES
)
1776 if (!node_state(node
, N_MEMORY
))
1780 if (!node_isset(node
, task_nodes
))
1783 if (current_node
== NUMA_NO_NODE
) {
1784 current_node
= node
;
1786 } else if (node
!= current_node
) {
1787 err
= move_pages_and_store_status(mm
, current_node
,
1788 &pagelist
, status
, start
, i
, nr_pages
);
1792 current_node
= node
;
1796 * Errors in the page lookup or isolation are not fatal and we simply
1797 * report them via status
1799 err
= add_page_for_migration(mm
, addr
, current_node
,
1800 &pagelist
, flags
& MPOL_MF_MOVE_ALL
);
1803 /* The page is successfully queued for migration */
1808 * The move_pages() man page does not have an -EEXIST choice, so
1809 * use -EFAULT instead.
1815 * If the page is already on the target node (!err), store the
1816 * node, otherwise, store the err.
1818 err
= store_status(status
, i
, err
? : current_node
, 1);
1822 err
= move_pages_and_store_status(mm
, current_node
, &pagelist
,
1823 status
, start
, i
, nr_pages
);
1826 current_node
= NUMA_NO_NODE
;
1829 /* Make sure we do not overwrite the existing error */
1830 err1
= move_pages_and_store_status(mm
, current_node
, &pagelist
,
1831 status
, start
, i
, nr_pages
);
1840 * Determine the nodes of an array of pages and store it in an array of status.
1842 static void do_pages_stat_array(struct mm_struct
*mm
, unsigned long nr_pages
,
1843 const void __user
**pages
, int *status
)
1849 for (i
= 0; i
< nr_pages
; i
++) {
1850 unsigned long addr
= (unsigned long)(*pages
);
1851 struct vm_area_struct
*vma
;
1855 vma
= vma_lookup(mm
, addr
);
1859 /* FOLL_DUMP to ignore special (like zero) pages */
1860 page
= follow_page(vma
, addr
, FOLL_GET
| FOLL_DUMP
);
1862 err
= PTR_ERR(page
);
1866 if (page
&& !is_zone_device_page(page
)) {
1867 err
= page_to_nid(page
);
1879 mmap_read_unlock(mm
);
1882 static int get_compat_pages_array(const void __user
*chunk_pages
[],
1883 const void __user
* __user
*pages
,
1884 unsigned long chunk_nr
)
1886 compat_uptr_t __user
*pages32
= (compat_uptr_t __user
*)pages
;
1890 for (i
= 0; i
< chunk_nr
; i
++) {
1891 if (get_user(p
, pages32
+ i
))
1893 chunk_pages
[i
] = compat_ptr(p
);
1900 * Determine the nodes of a user array of pages and store it in
1901 * a user array of status.
1903 static int do_pages_stat(struct mm_struct
*mm
, unsigned long nr_pages
,
1904 const void __user
* __user
*pages
,
1907 #define DO_PAGES_STAT_CHUNK_NR 16UL
1908 const void __user
*chunk_pages
[DO_PAGES_STAT_CHUNK_NR
];
1909 int chunk_status
[DO_PAGES_STAT_CHUNK_NR
];
1912 unsigned long chunk_nr
= min(nr_pages
, DO_PAGES_STAT_CHUNK_NR
);
1914 if (in_compat_syscall()) {
1915 if (get_compat_pages_array(chunk_pages
, pages
,
1919 if (copy_from_user(chunk_pages
, pages
,
1920 chunk_nr
* sizeof(*chunk_pages
)))
1924 do_pages_stat_array(mm
, chunk_nr
, chunk_pages
, chunk_status
);
1926 if (copy_to_user(status
, chunk_status
, chunk_nr
* sizeof(*status
)))
1931 nr_pages
-= chunk_nr
;
1933 return nr_pages
? -EFAULT
: 0;
1936 static struct mm_struct
*find_mm_struct(pid_t pid
, nodemask_t
*mem_nodes
)
1938 struct task_struct
*task
;
1939 struct mm_struct
*mm
;
1942 * There is no need to check if current process has the right to modify
1943 * the specified process when they are same.
1947 *mem_nodes
= cpuset_mems_allowed(current
);
1951 /* Find the mm_struct */
1953 task
= find_task_by_vpid(pid
);
1956 return ERR_PTR(-ESRCH
);
1958 get_task_struct(task
);
1961 * Check if this process has the right to modify the specified
1962 * process. Use the regular "ptrace_may_access()" checks.
1964 if (!ptrace_may_access(task
, PTRACE_MODE_READ_REALCREDS
)) {
1966 mm
= ERR_PTR(-EPERM
);
1971 mm
= ERR_PTR(security_task_movememory(task
));
1974 *mem_nodes
= cpuset_mems_allowed(task
);
1975 mm
= get_task_mm(task
);
1977 put_task_struct(task
);
1979 mm
= ERR_PTR(-EINVAL
);
1984 * Move a list of pages in the address space of the currently executing
1987 static int kernel_move_pages(pid_t pid
, unsigned long nr_pages
,
1988 const void __user
* __user
*pages
,
1989 const int __user
*nodes
,
1990 int __user
*status
, int flags
)
1992 struct mm_struct
*mm
;
1994 nodemask_t task_nodes
;
1997 if (flags
& ~(MPOL_MF_MOVE
|MPOL_MF_MOVE_ALL
))
2000 if ((flags
& MPOL_MF_MOVE_ALL
) && !capable(CAP_SYS_NICE
))
2003 mm
= find_mm_struct(pid
, &task_nodes
);
2008 err
= do_pages_move(mm
, task_nodes
, nr_pages
, pages
,
2009 nodes
, status
, flags
);
2011 err
= do_pages_stat(mm
, nr_pages
, pages
, status
);
2017 SYSCALL_DEFINE6(move_pages
, pid_t
, pid
, unsigned long, nr_pages
,
2018 const void __user
* __user
*, pages
,
2019 const int __user
*, nodes
,
2020 int __user
*, status
, int, flags
)
2022 return kernel_move_pages(pid
, nr_pages
, pages
, nodes
, status
, flags
);
2025 #ifdef CONFIG_NUMA_BALANCING
2027 * Returns true if this is a safe migration target node for misplaced NUMA
2028 * pages. Currently it only checks the watermarks which is crude.
2030 static bool migrate_balanced_pgdat(struct pglist_data
*pgdat
,
2031 unsigned long nr_migrate_pages
)
2035 for (z
= pgdat
->nr_zones
- 1; z
>= 0; z
--) {
2036 struct zone
*zone
= pgdat
->node_zones
+ z
;
2038 if (!managed_zone(zone
))
2041 /* Avoid waking kswapd by allocating pages_to_migrate pages. */
2042 if (!zone_watermark_ok(zone
, 0,
2043 high_wmark_pages(zone
) +
2052 static struct page
*alloc_misplaced_dst_page(struct page
*page
,
2055 int nid
= (int) data
;
2056 int order
= compound_order(page
);
2057 gfp_t gfp
= __GFP_THISNODE
;
2061 gfp
|= GFP_TRANSHUGE_LIGHT
;
2063 gfp
|= GFP_HIGHUSER_MOVABLE
| __GFP_NOMEMALLOC
| __GFP_NORETRY
|
2065 gfp
&= ~__GFP_RECLAIM
;
2067 new = __folio_alloc_node(gfp
, order
, nid
);
2072 static int numamigrate_isolate_page(pg_data_t
*pgdat
, struct page
*page
)
2074 int nr_pages
= thp_nr_pages(page
);
2075 int order
= compound_order(page
);
2077 VM_BUG_ON_PAGE(order
&& !PageTransHuge(page
), page
);
2079 /* Do not migrate THP mapped by multiple processes */
2080 if (PageTransHuge(page
) && total_mapcount(page
) > 1)
2083 /* Avoid migrating to a node that is nearly full */
2084 if (!migrate_balanced_pgdat(pgdat
, nr_pages
)) {
2087 if (!(sysctl_numa_balancing_mode
& NUMA_BALANCING_MEMORY_TIERING
))
2089 for (z
= pgdat
->nr_zones
- 1; z
>= 0; z
--) {
2090 if (managed_zone(pgdat
->node_zones
+ z
))
2093 wakeup_kswapd(pgdat
->node_zones
+ z
, 0, order
, ZONE_MOVABLE
);
2097 if (isolate_lru_page(page
))
2100 mod_node_page_state(page_pgdat(page
), NR_ISOLATED_ANON
+ page_is_file_lru(page
),
2104 * Isolating the page has taken another reference, so the
2105 * caller's reference can be safely dropped without the page
2106 * disappearing underneath us during migration.
2113 * Attempt to migrate a misplaced page to the specified destination
2114 * node. Caller is expected to have an elevated reference count on
2115 * the page that will be dropped by this function before returning.
2117 int migrate_misplaced_page(struct page
*page
, struct vm_area_struct
*vma
,
2120 pg_data_t
*pgdat
= NODE_DATA(node
);
2123 unsigned int nr_succeeded
;
2124 LIST_HEAD(migratepages
);
2125 int nr_pages
= thp_nr_pages(page
);
2128 * Don't migrate file pages that are mapped in multiple processes
2129 * with execute permissions as they are probably shared libraries.
2131 if (page_mapcount(page
) != 1 && page_is_file_lru(page
) &&
2132 (vma
->vm_flags
& VM_EXEC
))
2136 * Also do not migrate dirty pages as not all filesystems can move
2137 * dirty pages in MIGRATE_ASYNC mode which is a waste of cycles.
2139 if (page_is_file_lru(page
) && PageDirty(page
))
2142 isolated
= numamigrate_isolate_page(pgdat
, page
);
2146 list_add(&page
->lru
, &migratepages
);
2147 nr_remaining
= migrate_pages(&migratepages
, alloc_misplaced_dst_page
,
2148 NULL
, node
, MIGRATE_ASYNC
,
2149 MR_NUMA_MISPLACED
, &nr_succeeded
);
2151 if (!list_empty(&migratepages
)) {
2152 list_del(&page
->lru
);
2153 mod_node_page_state(page_pgdat(page
), NR_ISOLATED_ANON
+
2154 page_is_file_lru(page
), -nr_pages
);
2155 putback_lru_page(page
);
2160 count_vm_numa_events(NUMA_PAGE_MIGRATE
, nr_succeeded
);
2161 if (!node_is_toptier(page_to_nid(page
)) && node_is_toptier(node
))
2162 mod_node_page_state(pgdat
, PGPROMOTE_SUCCESS
,
2165 BUG_ON(!list_empty(&migratepages
));
2172 #endif /* CONFIG_NUMA_BALANCING */
2175 * node_demotion[] example:
2177 * Consider a system with two sockets. Each socket has
2178 * three classes of memory attached: fast, medium and slow.
2179 * Each memory class is placed in its own NUMA node. The
2180 * CPUs are placed in the node with the "fast" memory. The
2181 * 6 NUMA nodes (0-5) might be split among the sockets like
2187 * When Node 0 fills up, its memory should be migrated to
2188 * Node 1. When Node 1 fills up, it should be migrated to
2189 * Node 2. The migration path start on the nodes with the
2190 * processors (since allocations default to this node) and
2191 * fast memory, progress through medium and end with the
2194 * 0 -> 1 -> 2 -> stop
2195 * 3 -> 4 -> 5 -> stop
2197 * This is represented in the node_demotion[] like this:
2199 * { nr=1, nodes[0]=1 }, // Node 0 migrates to 1
2200 * { nr=1, nodes[0]=2 }, // Node 1 migrates to 2
2201 * { nr=0, nodes[0]=-1 }, // Node 2 does not migrate
2202 * { nr=1, nodes[0]=4 }, // Node 3 migrates to 4
2203 * { nr=1, nodes[0]=5 }, // Node 4 migrates to 5
2204 * { nr=0, nodes[0]=-1 }, // Node 5 does not migrate
2206 * Moreover some systems may have multiple slow memory nodes.
2207 * Suppose a system has one socket with 3 memory nodes, node 0
2208 * is fast memory type, and node 1/2 both are slow memory
2209 * type, and the distance between fast memory node and slow
2210 * memory node is same. So the migration path should be:
2214 * This is represented in the node_demotion[] like this:
2215 * { nr=2, {nodes[0]=1, nodes[1]=2} }, // Node 0 migrates to node 1 and node 2
2216 * { nr=0, nodes[0]=-1, }, // Node 1 dose not migrate
2217 * { nr=0, nodes[0]=-1, }, // Node 2 does not migrate
2221 * Writes to this array occur without locking. Cycles are
2222 * not allowed: Node X demotes to Y which demotes to X...
2224 * If multiple reads are performed, a single rcu_read_lock()
2225 * must be held over all reads to ensure that no cycles are
2228 #define DEFAULT_DEMOTION_TARGET_NODES 15
2230 #if MAX_NUMNODES < DEFAULT_DEMOTION_TARGET_NODES
2231 #define DEMOTION_TARGET_NODES (MAX_NUMNODES - 1)
2233 #define DEMOTION_TARGET_NODES DEFAULT_DEMOTION_TARGET_NODES
2236 struct demotion_nodes
{
2238 short nodes
[DEMOTION_TARGET_NODES
];
2241 static struct demotion_nodes
*node_demotion __read_mostly
;
2244 * next_demotion_node() - Get the next node in the demotion path
2245 * @node: The starting node to lookup the next node
2247 * Return: node id for next memory node in the demotion path hierarchy
2248 * from @node; NUMA_NO_NODE if @node is terminal. This does not keep
2249 * @node online or guarantee that it *continues* to be the next demotion
2252 int next_demotion_node(int node
)
2254 struct demotion_nodes
*nd
;
2255 unsigned short target_nr
, index
;
2259 return NUMA_NO_NODE
;
2261 nd
= &node_demotion
[node
];
2264 * node_demotion[] is updated without excluding this
2265 * function from running. RCU doesn't provide any
2266 * compiler barriers, so the READ_ONCE() is required
2267 * to avoid compiler reordering or read merging.
2269 * Make sure to use RCU over entire code blocks if
2270 * node_demotion[] reads need to be consistent.
2273 target_nr
= READ_ONCE(nd
->nr
);
2275 switch (target_nr
) {
2277 target
= NUMA_NO_NODE
;
2284 * If there are multiple target nodes, just select one
2285 * target node randomly.
2287 * In addition, we can also use round-robin to select
2288 * target node, but we should introduce another variable
2289 * for node_demotion[] to record last selected target node,
2290 * that may cause cache ping-pong due to the changing of
2291 * last target node. Or introducing per-cpu data to avoid
2292 * caching issue, which seems more complicated. So selecting
2293 * target node randomly seems better until now.
2295 index
= get_random_int() % target_nr
;
2299 target
= READ_ONCE(nd
->nodes
[index
]);
2306 /* Disable reclaim-based migration. */
2307 static void __disable_all_migrate_targets(void)
2314 for_each_online_node(node
) {
2315 node_demotion
[node
].nr
= 0;
2316 for (i
= 0; i
< DEMOTION_TARGET_NODES
; i
++)
2317 node_demotion
[node
].nodes
[i
] = NUMA_NO_NODE
;
2321 static void disable_all_migrate_targets(void)
2323 __disable_all_migrate_targets();
2326 * Ensure that the "disable" is visible across the system.
2327 * Readers will see either a combination of before+disable
2328 * state or disable+after. They will never see before and
2329 * after state together.
2331 * The before+after state together might have cycles and
2332 * could cause readers to do things like loop until this
2333 * function finishes. This ensures they can only see a
2334 * single "bad" read and would, for instance, only loop
2341 * Find an automatic demotion target for 'node'.
2342 * Failing here is OK. It might just indicate
2343 * being at the end of a chain.
2345 static int establish_migrate_target(int node
, nodemask_t
*used
,
2348 int migration_target
, index
, val
;
2349 struct demotion_nodes
*nd
;
2352 return NUMA_NO_NODE
;
2354 nd
= &node_demotion
[node
];
2356 migration_target
= find_next_best_node(node
, used
);
2357 if (migration_target
== NUMA_NO_NODE
)
2358 return NUMA_NO_NODE
;
2361 * If the node has been set a migration target node before,
2362 * which means it's the best distance between them. Still
2363 * check if this node can be demoted to other target nodes
2364 * if they have a same best distance.
2366 if (best_distance
!= -1) {
2367 val
= node_distance(node
, migration_target
);
2368 if (val
> best_distance
)
2373 if (WARN_ONCE(index
>= DEMOTION_TARGET_NODES
,
2374 "Exceeds maximum demotion target nodes\n"))
2377 nd
->nodes
[index
] = migration_target
;
2380 return migration_target
;
2382 node_clear(migration_target
, *used
);
2383 return NUMA_NO_NODE
;
2387 * When memory fills up on a node, memory contents can be
2388 * automatically migrated to another node instead of
2389 * discarded at reclaim.
2391 * Establish a "migration path" which will start at nodes
2392 * with CPUs and will follow the priorities used to build the
2393 * page allocator zonelists.
2395 * The difference here is that cycles must be avoided. If
2396 * node0 migrates to node1, then neither node1, nor anything
2397 * node1 migrates to can migrate to node0. Also one node can
2398 * be migrated to multiple nodes if the target nodes all have
2399 * a same best-distance against the source node.
2401 * This function can run simultaneously with readers of
2402 * node_demotion[]. However, it can not run simultaneously
2403 * with itself. Exclusion is provided by memory hotplug events
2404 * being single-threaded.
2406 static void __set_migration_target_nodes(void)
2408 nodemask_t next_pass
;
2409 nodemask_t this_pass
;
2410 nodemask_t used_targets
= NODE_MASK_NONE
;
2411 int node
, best_distance
;
2414 * Avoid any oddities like cycles that could occur
2415 * from changes in the topology. This will leave
2416 * a momentary gap when migration is disabled.
2418 disable_all_migrate_targets();
2421 * Allocations go close to CPUs, first. Assume that
2422 * the migration path starts at the nodes with CPUs.
2424 next_pass
= node_states
[N_CPU
];
2426 this_pass
= next_pass
;
2427 next_pass
= NODE_MASK_NONE
;
2429 * To avoid cycles in the migration "graph", ensure
2430 * that migration sources are not future targets by
2431 * setting them in 'used_targets'. Do this only
2432 * once per pass so that multiple source nodes can
2433 * share a target node.
2435 * 'used_targets' will become unavailable in future
2436 * passes. This limits some opportunities for
2437 * multiple source nodes to share a destination.
2439 nodes_or(used_targets
, used_targets
, this_pass
);
2441 for_each_node_mask(node
, this_pass
) {
2445 * Try to set up the migration path for the node, and the target
2446 * migration nodes can be multiple, so doing a loop to find all
2447 * the target nodes if they all have a best node distance.
2451 establish_migrate_target(node
, &used_targets
,
2454 if (target_node
== NUMA_NO_NODE
)
2457 if (best_distance
== -1)
2458 best_distance
= node_distance(node
, target_node
);
2461 * Visit targets from this pass in the next pass.
2462 * Eventually, every node will have been part of
2463 * a pass, and will become set in 'used_targets'.
2465 node_set(target_node
, next_pass
);
2469 * 'next_pass' contains nodes which became migration
2470 * targets in this pass. Make additional passes until
2471 * no more migrations targets are available.
2473 if (!nodes_empty(next_pass
))
2478 * For callers that do not hold get_online_mems() already.
2480 void set_migration_target_nodes(void)
2483 __set_migration_target_nodes();
2488 * This leaves migrate-on-reclaim transiently disabled between
2489 * the MEM_GOING_OFFLINE and MEM_OFFLINE events. This runs
2490 * whether reclaim-based migration is enabled or not, which
2491 * ensures that the user can turn reclaim-based migration at
2492 * any time without needing to recalculate migration targets.
2494 * These callbacks already hold get_online_mems(). That is why
2495 * __set_migration_target_nodes() can be used as opposed to
2496 * set_migration_target_nodes().
2498 #ifdef CONFIG_MEMORY_HOTPLUG
2499 static int __meminit
migrate_on_reclaim_callback(struct notifier_block
*self
,
2500 unsigned long action
, void *_arg
)
2502 struct memory_notify
*arg
= _arg
;
2505 * Only update the node migration order when a node is
2506 * changing status, like online->offline. This avoids
2507 * the overhead of synchronize_rcu() in most cases.
2509 if (arg
->status_change_nid
< 0)
2510 return notifier_from_errno(0);
2513 case MEM_GOING_OFFLINE
:
2515 * Make sure there are not transient states where
2516 * an offline node is a migration target. This
2517 * will leave migration disabled until the offline
2518 * completes and the MEM_OFFLINE case below runs.
2520 disable_all_migrate_targets();
2525 * Recalculate the target nodes once the node
2526 * reaches its final state (online or offline).
2528 __set_migration_target_nodes();
2530 case MEM_CANCEL_OFFLINE
:
2532 * MEM_GOING_OFFLINE disabled all the migration
2533 * targets. Reenable them.
2535 __set_migration_target_nodes();
2537 case MEM_GOING_ONLINE
:
2538 case MEM_CANCEL_ONLINE
:
2542 return notifier_from_errno(0);
2546 void __init
migrate_on_reclaim_init(void)
2548 node_demotion
= kcalloc(nr_node_ids
,
2549 sizeof(struct demotion_nodes
),
2551 WARN_ON(!node_demotion
);
2552 #ifdef CONFIG_MEMORY_HOTPLUG
2553 hotplug_memory_notifier(migrate_on_reclaim_callback
, 100);
2556 * At this point, all numa nodes with memory/CPus have their state
2557 * properly set, so we can build the demotion order now.
2558 * Let us hold the cpu_hotplug lock just, as we could possibily have
2559 * CPU hotplug events during boot.
2562 set_migration_target_nodes();
2566 bool numa_demotion_enabled
= false;
2569 static ssize_t
numa_demotion_enabled_show(struct kobject
*kobj
,
2570 struct kobj_attribute
*attr
, char *buf
)
2572 return sysfs_emit(buf
, "%s\n",
2573 numa_demotion_enabled
? "true" : "false");
2576 static ssize_t
numa_demotion_enabled_store(struct kobject
*kobj
,
2577 struct kobj_attribute
*attr
,
2578 const char *buf
, size_t count
)
2582 ret
= kstrtobool(buf
, &numa_demotion_enabled
);
2589 static struct kobj_attribute numa_demotion_enabled_attr
=
2590 __ATTR(demotion_enabled
, 0644, numa_demotion_enabled_show
,
2591 numa_demotion_enabled_store
);
2593 static struct attribute
*numa_attrs
[] = {
2594 &numa_demotion_enabled_attr
.attr
,
2598 static const struct attribute_group numa_attr_group
= {
2599 .attrs
= numa_attrs
,
2602 static int __init
numa_init_sysfs(void)
2605 struct kobject
*numa_kobj
;
2607 numa_kobj
= kobject_create_and_add("numa", mm_kobj
);
2609 pr_err("failed to create numa kobject\n");
2612 err
= sysfs_create_group(numa_kobj
, &numa_attr_group
);
2614 pr_err("failed to register numa group\n");
2620 kobject_put(numa_kobj
);
2623 subsys_initcall(numa_init_sysfs
);
2624 #endif /* CONFIG_SYSFS */
2625 #endif /* CONFIG_NUMA */